Research On Optimization Of Abnormal Behavior Detection Model And Method In Surveillance Video

At present,video surveillance system has been widely used in all kinds of public places,in the protection of social security,fighting crime,ensuring people’s lives and other aspects of play an increasingly critical role.With the rapid development of big data and artificial intelligence technology,intelligent video surveillance system with intelligent video information analysis function has become an urgent problem to be solved in the field of public security.It has important research significance and practical value for promoting the construction of "smart city" and "safe city",improving people’s sense of security and urban public security management and service ability.Therefore,based on the existing abnormal behavior detection model and deep learning method,this paper aims at the improvement of abnormal behavior detection model and method optimization in surveillance videos in public places.The main research work is as follows.1.An optimized video abnormal behavior detection model(MCS-YOLO)based on improved YOLOv5 network is proposed.In view of the problems that the detection accuracy is difficult to improve due to the large interference of background information in the monitoring video environment in public places and different target scales of abnormal behaviors in the video,an improved abnormal behavior detection model of YOLOv5 network is designed.MCS-YOLO adds a masked convolution attention model to the YOLOv5 backbone network.The model starts from a masked convolution layer,and the central region of the receptor field is masked.The model predicts the masked information and uses the errors related to the masked information as the anomaly score.At the same time,the Swin-CA module is embedded in the detection network.By learning the features of adjacent layers,the model can better grasp the global information,thus reducing the influence of background information on the detection results.By extracting the scale features of abnormal behavior in different backgrounds,the computational complexity of the whole model is reduced,and the positioning accuracy of the target with abnormal behavior in the video is improved.Experimental results on UCSD-ped1,KTH and Shanghai Tech data sets show that the accuracy of the proposed model reaches 98.2%,96.4% and 95.8%,respectively.2.A video abnormal behavior detection network based on reconstruction and prediction is proposed.In order to make full use of action and spatiotemporal characteristic information in abnormal behavior detection,an abnormal behavior detection network is designed which can obtain both action and spatiotemporal information.The network is composed of a reconstruction subnetwork and a video prediction subnetwork.The reconstruction subnetwork uses an autoencoder structure to reconstruct the next frame with continuous video frames as input,while the prediction subnetwork uses an encoder and decoder structure based on 3D convolution as the network backbone to predict subsequent video frames through the input of a series of video frames.In addition,in order to make the reconstructed subnetwork pay more attention to the action characteristics of the abnormal behavior in the video,Jansen-Shannon divergence(JSD)is used to calculate the difference between the reconstructed frame and the original frame.Meanwhile,the regularization constraint of spatio-temporal consistency is added to the prediction subnetwork to obtain the spatio-temporal characteristics of the abnormal behavior by using the time continuity of the action in the video.Experiments on UCSDped2,Avenue and Shanghai Tech data sets show that the proposed method has better performance than other abnormal behavior detection methods,and the average area under the curve of the three data sets reaches98.5%,92.1% and 87.6%.3.A video anomaly detection method based on optimized prediction network is proposed.Most of the existing video abnormal behavior detection methods are mainly based on the minimum reconstruction error criterion,and lack of consideration of the longterm abnormal behavior motion.Therefore,a prediction network based on motion context awareness is proposed to detect video abnormal behavior.This method introduces the longterm motion context memory with memory alignment learning,which can store the information of video frames into memory,establish the mapping relationship between the current input frame and the previous frame to facilitate the subsequent prediction,and detect the abnormal behavior of video targets by predicting the difference between the future frame and the original frame.At the same time,the global matching optical flow estimation method with overlapping attention is used to obtain the optical flow graphs of the future frame and the original frame,and the motion constraint is introduced in the prediction to keep the optical flow between the future frame and the original frame consistent.In addition,the generative adversarial network is used to train the model,so that the method can better predict the future frame and improve the accuracy of abnormal behavior detection.The experimental results on UCSDped2,Avenue and Shanghai Tech data sets show that the proposed method has better performance than other human abnormal behavior detection methods,reaching 96.7%,90.7% and 85.4% in UCSDped2,Avenue and Shanghai Tech data sets,respectively.In this paper,the existing video abnormal behavior detection methods are studied in depth,focusing on the structural improvement and optimization of YOLOv5 network,reconstruction network and prediction network,so that the accuracy of the improved method and model in the video abnormal detection task has been greatly improved,laying a technical foundation for the further development of intelligent surveillance video system.